Epoc BCI scenario

Working with OpenViBE signal processing scenarios and doing scenario/BCI design
Post Reply
morozov
Posts: 5
Joined: Tue Jun 17, 2014 2:19 pm

Epoc BCI scenario

Post by morozov »

Hello!

I'm trying to build an interactive application using the Emotiv Epoc as an EEG.
The only thing left to do is the signal processing.

- and as I'm baffled by the terminology and lost in the existing scenarios this is my call for help.

I'm simulating sound sources in a virtual space, with which the user can interact (realtime)-

I need to classify the thought of "pulling" a sound source towards oneself,
in order to make the sound source come closer.

Could someone please point me in the right direction?
Like, should I use a Spatial Filter with this scenario, and, if yes, how do I find some appropriate settings?
Is there a scenario that does this kind of thing?

I'm a BCI-newb, sorry if these questions have been answered,
my searches leave me disoriented.

This is what I have:
eegamma.zip
This contains my attempt
(70.96 KiB) Downloaded 242 times
Thanks a bunch.

Kind regards,
Alexander Morosow
Interaction Design

jtlindgren
Posts: 775
Joined: Tue Dec 04, 2012 3:53 pm
Location: INRIA Rennes, FRANCE

Re: Epoc BCI scenario

Post by jtlindgren »

Hi Morozov,

I'm not sure if the thought of 'pulling' can be reliably detected by EEG, its definitely not one of the main paradigms.

The way you can try this is to first collect a set of data (see the motor imagery scenarios) where for some segments
of time the user is instructed to think of pulling, and for others, to relax (for example). These time segments of EEG
data are then turned into 'features' which are then used to train a classifier along with the knowledge what the
user was supposedly thinking (the 'correct classifications' for each segment, or 'labels'). This classifier can be later
used to try to distinguish between pulling and not pulling. The output controls your application.

Fabien Lotte's bibliography has good pointers to get you started,

https://sites.google.com/site/fabienlotte/bibliography

You might look at the overview or motor imagery papers. It is possible that pulling might not be easy to detect
and the pulling is easier to base on detecting some other mental activity.


My three cents,
Jussi

fabien.lotte
Posts: 112
Joined: Sun Mar 14, 2010 12:58 pm

Re: Epoc BCI scenario

Post by fabien.lotte »

Dear Morozov,

To complement what Jussi has said, you can also look at the videolectures website, there are a number of tutorials on EEG signal processing and classification:

http://videolectures.net/Top/Technology ... echnology/

As Jussi said, you can start by using the motor imagery scenario from openvibe, that would be a good start. But Indeed, recognizing imagining pulling a sound from EEG signals is probably not something you can easily to do. I know Emotiv claim they can do it, but it has never been formally validated (to the best of my knowledge there is not a single scientific paper showing that it actually works), so I wouldn't expect too much from that.

I hope this helps,

best,
Fabien

morozov
Posts: 5
Joined: Tue Jun 17, 2014 2:19 pm

Re: Epoc BCI scenario

Post by morozov »

Hello, thank you both for the feedback, advice and resources.
That helped me a lot in understanding the caveats of my project.

After experimenting a bit with the signal processing I am able to achieve a classifier performance of 100%.
Does this number mean anything in respect to the real world performance?

When I run the same data through my replay scenario, I get a miserable (8-15%) classifier accuracy -

I'm using the LDA classifier and CSP spatial filter.
The signal processing in the classification and the replay scenario is identical, as is the EEG data for now.

Can anybody tell why this could be the case?

Thanks a lot!

Kind regards,
Alexander

Edit:
I suppose that some of the causes can be due to insufficient a-priori knowledge about the patterns involved in thinking about "pulling", so that it is hard to grasp the relevant information.
To my understanding, I should get reasonable accuracy if I use the same data in classifier training and "replay", though.
Please correct me if I'm wrong on that!

jtlindgren
Posts: 775
Joined: Tue Dec 04, 2012 3:53 pm
Location: INRIA Rennes, FRANCE

Re: Epoc BCI scenario

Post by jtlindgren »

Your classifier might be overfitting the data. You could try using Shrinkage LDA or the SVM and make the regularization heavier until you reach lower training accuracies, like 80% and see if the test error starts to resemble the training error more.

If the data is really identical then you should get very similar accuracies in both train and test. If the accuracies are very different, there is a bug somewhere.


Happy hacking,
Jussi

Post Reply